Jira
SlackWhat they did
David Abitbol, who leads AI operations and infrastructure at Revolut, described the rollout at the AI Engineer conference in New York in October 2026. Revolut tried Gemini in early 2025, then Claude Code at scale in early 2026 with strict per-engineer budgets, because some engineers were spending $1,000 to $10,000 a month. In April 2026 it began hosting Z.ai's open-weight GLM 5.1 itself, later GLM 5.3, on 56 Nvidia B200 GPUs, and made it the default coding agent for about 2,000 engineers. A router sends image and UI work to commercial models, since GLM is text only. With no token caps, engineers also use it for debugging and for working through Jira and Confluence tickets.
How it went
Blended cost is about $0.30 per million tokens, roughly a tenth of what Revolut paid Anthropic and OpenAI, on a flat bill. Revolut measured coding time and lead time from pickup to merge: about 20% and 10% better with Gemini, little change under the capped Claude Code rollout, then 60% and 70% better after the switch to GLM. Tuning, NVFP4 quantization, split prefill and decode, and a 256K context cap got first tokens under three seconds at about 200 tokens a second. Claude Code and Codex stay available as exceptions.
Worth knowing
Commenters doubted the 60% figure, pointing out it may compare unlimited GLM against budget-capped Claude rather than like for like, and that it comes from the project's own lead. Abitbol said first-token latency over five seconds fills his Slack with complaints, and that a million-token context is not worth serving. The harness engineers run GLM in was not named. The hardware is a seven-figure outlay.